Match AniGen SLat DAE head FP32 output contract

Expand fixed-cardinality SLat DAE head profiles to 8192 tokens

Set L4 SLat DAE x64 head max profile to 4096 tokens

Use unique external-data files per SLat DAE ONNX branch

Bound x64 SLat DAE head TensorRT profiles for L4 memory

Use static normalized_shape in SLat DAE ONNX LayerNorm export

Parse SLat DAE ONNX from file so TensorRT resolves external data

Add concise failure diagnostics around SLat DAE branch ONNX export

Add TensorRT builder for SLat DAE custom ONNX branches and heads

Accept spconv list kernel_size in SLat DAE custom head export

Use torch.library sparse custom ops and native wrapped spconv weights in SLat DAE heads

Use torch.library custom op for stable SLat DAE ONNX window attention export

Add SLat DAE custom ONNX upsample and skeleton head exporter

Match AniGen FP32 input projections before FP16 SLat DAE blocks

Add full SLat DAE Swin branch ONNX exporter using custom sparse attention

Expand AniGen skin TensorRT profile to 300k vertices

Add AniGen SLat sparse-shell checkpoint extractor

Add AniGen skin-decoder TensorRT profile

Add AniGen skin decoder ONNX exporter

Add dynamic profiles for AniGen SLat TensorRT cores

Add AniGen SLat transformer core ONNX exporter

Add TensorRT builder for AniGen ONNX components

Add full AniGen SS decoder ONNX exporter

Add full AniGen SS Flow ONNX exporter

Add static FLUX.2 Klein TensorRT L4 builder

Use streaming ONNX checker for multi-GB FLUX.2 export

Fix FLUX.2 external-data ONNX validation

Split TensorRT builds per component to stay within 30GB RAM

Use torch.export ONNX opset 23 for FLUX.2 RMSNormalization

Use NVIDIA ModelOpt AutoCast for TensorRT 11 mixed FP16

Export FLUX.2 transformer on L4 CUDA with static profiles

Tune TensorRT L4 builder optimization level

Support TensorRT 10 explicit-batch API

Add static FLUX.2 Klein transformer ONNX export pipeline

Fix DSINE ONNX export on CUDA

Add L4 ONNX TensorRT hybrid runtime scaffold